Zhijing Zhang, Jinpeng Yu, Xin Song, Bingnan Li, Chuyue Li, Changhui Du, Xiaolin Fang, Jiaming Liu, Ruihua Huang
5 min
Abstract
Conversational assistants increasingly recommend follow-up edits to help users continue a task. Existing systems primarily target text-only interactions, leaving image-creation conversations underexplored. In image-creation tasks, useful follow-up edit suggestions must reflect user preferences, offer diverse directions, and remain executable on the current image. We collected 100,000 real multi-turn image-creation conversation samples from Qwen App and found that 80.1% are image-dependent, underscoring the need for multimodal recommendation. We address this setting with a three-stage framework. In Stage 1, we use real online data to build a human-reviewed table of appropriate follow-up editing intents, then create SFT targets and fine-tune a multimodal policy. In Stage 2, to align rule-guided SFT suggestions with actual user choices, we use user click feedback to optimize the policy through multi-objective reinforcement learning. In Stage 3, to reduce visual inconsistencies between suggested edits and the current image, we introduce a visual verifier as additional training supervision. Extensive experiments demonstrate that our framework significantly outperforms baselines on both automatic and human evaluations. In a live user-randomized A/B test with millions of users, our final framework reduces visual inconsistency from 3.7% to 0.9%. Furthermore, it significantly improves recommendation CTR by 32.70%, image take-away rate by 16.32%, and average conversation turns per user by 39.90% (all p<0.05).
Alex: That's a clever fix. But if clicks already tell you what people like, why do you need a third stage at all?
Sam: Because "what people click" and "what actually fits the image" are not the same thing. A suggestion can sound appealing and still be completely wrong for the picture. For example, it might suggest adding lens flare to a photo that was taken indoors with no light source to create flare from. Sounds cool, doesn't fit. So the third stage is an image-first check.
Alex: What does that check actually do?
Sam: Think of it like a careful editor who reads the picture before reading the suggestion. It asks two questions: first, does the suggestion depend on something that isn't actually in the image? And second, is the change being requested something that's already true? If the answer to either is yes, the suggestion gets flagged. The paper calls this a source-target structure — what must already exist, and what is supposed to change — but the everyday idea is simple: does this edit make sense for this specific picture?
Alex: Why split it into two questions instead of one overall judgment?
Sam: Because different kinds of mistakes need different checks. "Missing starting point" and "already done" are two separate failure modes. A broad judgment might miss one of them. By separating the checks, the system catches both more reliably.
Alex: What if the image is ambiguous — like, it's not obvious whether something is present or not?
Sam: The paper is careful about that. If the image doesn't give enough evidence to decide clearly, the check fails open rather than forcing a judgment. So the system is strict when the image is clear, and cautious when it isn't. That keeps it from pretending to know more than it does.
Alex: So the preference stage makes suggestions more appealing, but can also make them drift — more repetitive, more likely to clash with the image — and the visual check pulls them back?
Sam: That's exactly the tradeoff the paper identifies. The preference stage improves what expert reviewers judge as quality, but it introduces those side effects. The visual check acts as a guardrail. And when they tested the full system in Qwen App — a real deployment with real users — they saw improvements in how often people clicked suggestions, how often they actually applied edits, and how long they kept editing. The paper suggests that happened because the system was no longer choosing between being appealing and being correct. It was pushed toward both at once.
Alex: So the core lesson is that a good suggestion has to be liked, but it also has to make sense for the actual image in front of you.
Sam: Right. And that sounds straightforward, but building a system that reliably does both — at scale, in a real app — turns out to require all three stages working together. Human review to establish the task, click feedback to learn user preference, and an image check to keep the system grounded in visual reality. Remove any one of those, and something breaks.
Alex: That's a genuinely useful framing — not just for photo editing, but for any AI system that has to give advice about something visual.
Sam: That's a fair takeaway. The paper is specifically about image editing suggestions, but the underlying problem — how do you make a recommendation that's both appealing and actually valid for the specific context — shows up in a lot of places. Thanks for listening to ResearchPod.